AI Machine Translation: How Good Is It?

Opening

If you only speak one language and need your content translated, you might be wondering if you could use AI to do it. Machine translation software and generative AI are very good at translating nowadays and most of the time, free. Still, there are many professional human translators. Should you spend money on a translation agency? Should you find a machine translation post-editor? Should you just do it for free with ChatGPT or Google Translate?

My Background

I’ll give you an inside look as a translator, reader and writer of multiple languages. To give you a better understanding of where I am coming from, I studied philosophy of language and French language in university and have been providing English and Korean language services (translation, localisation, copywriting, copy editing, content marketing, voice-over) for nearly a decade.

Understand How Translations Are Done

Different Workflow for Different Contents

For human translators, there are various ways to translate different materials depending on the length, format, genre and goals of the content. It’s typical and expected from translators who work with document/research or software/app to use CAT* tools.

On the other hand, a good portion of translators who work on books rarely use them. Clients’ expectations also change based on the material. Document/software translators are expected to be timely and communicative while book translators are primarily expected to provide high quality writing in the target language.

* In a nutshell, a CAT tool cuts up the source text into meaningful units. Like, [In a nutshell], [a], [CAT tool] and so on. Then, it matches them with the data that the user had provided, so the user can either confirm it or edit it quickly. If you are translating from English to French, it will be like this: [In a nutshell] ⇒ [En bref].

The Essence of Translation

All translators are editors. Though not all types of translators need to be master writers, they all should know how to avoid writing bad sentences in their target language.

These days, there are many gigs titled “machine translation post-editing.” Calling the job “post-editing” makes it sound like the translator would contribute much less than in a regular translation, which therefore, seemingly justifies a lower pay rate. Although MT post-editing does take less time and allows translators to take on more projects per work day, translators’ effort and contribution don’t decrease pro rata to the lowered rate (or anywhere close to it, if the translator is quality oriented).

Why? The translators’ main expertise, where they create the most value has always been editing. Before AI, there were pattern matching MT and CAT software. Even before those, there were dictionaries. These tools were capable of providing potentially correct matches between the source and the target language.

From there, translators have been specialised to choose the best matches, using their cultural knowledge and language proficiency, to write up something entirely new when the pre-existing matches aren’t appropriate, to make sure sentences flow well with consistent style and finally, to verify the equivalence of the source text and the result. (The final part is quite significant in terms of fee; Translation agencies usually have three people to ensure this verification, namely translator, proofreader and project manager.)

For comparison, think of a lawyer. Instead of writing clients’ contracts, they can start “machine contract post-editing.” The lawyer can save some time letting their AI do the writing and formatting. Nonetheless, their responsibility and the fee would remain more or less the same. Because a lawyer’s expertise is not in making a contract but confirming that the contract is alright. Their speciality lies in applying their knowledge correctly on the client’s specific situation and needs.

Back to Content Strategy

To evaluate your needs in translation, go back to the beginning. What is the message you want to send to your audience through your content—whether it be a market report, video game, website, or a package label? What are your goals? Is it to tell your audience information, to appeal to their emotion, to express your world or to give them a certain experience? Maybe you aim to establish a persona or build rapport.

Perhaps the text of a fortune cookie could be done with machine translation only. Your brand story shouldn’t be. A brand story isn’t primarily about literal meaning, and probably also isn’t about making funny culturally inappropriate errors. Its power and value rely heavily on shared cultural context and linguistic intuitions.

There needs to be a sense of where your words are coming from and how and with whom they are meant to land, which AI isn’t built for. AI doesn’t interpret anything; It only guesses. AI processes the text and outputs very well-educated guesses.

Generally speaking, when your goal is to communicate something subtextual, such as character dynamic, creative expression or emotional persuasion—or when the cost of errors is high—your content probably needs a more human touch.

Your Language Pair Matters

Language pairs also make a big difference. For machine translation tools, training data is the most abundant in English. When the target language is English, the result will be the most fluent and well-written. When your source or target language is less accessible to people in Silicon Valley, not so much.

Even in the best possible scenario where your target language is English and the source language is one of the other major world languages, the result can vary. Structurally and lexically, German or Italian are much more similar to English than Japanese or Chinese even though they all are widely spoken.

For example, a Korean sentence can say something like “Eat.” it could mean any of these: “I eat”, “I am eating”, “She/he/they eat(s)”, “You eat [imperative]”. Many gender and number markers are absent in Korean and present in English. MT results show pronouns all over the place. Machine translation technology made a drastic improvement on picking up these contexts during the last 10 years or so. But it still makes mistakes. Here’s an example I found at my first attempt with Google Translate:

• My original text in Korean: “하늘이는 밥을 먹었다. 하지만 배가 고팠다.”
• Google Translate (Korean to English): “Sky ate rice. But I was hungry.”
• My translation: “Haneul had a meal. But she was hungry.”

Many language pairs, I believe, will always need some degree of human touch just to avoid objectively wrong translation. I recommend you to compare risk and cost to see how economic it would really be to remove a professional from the translation process and to risk having confusing mistakes in your specific content, and to consider what you are actually paying for when you pay for a human professional.

Let’s Revisit Your Goals

In day-to-day, immediate communications, like messaging friends or colleagues who don’t speak the same language as you, all you need is to get your point across. Your good intentions can be conveyed with non-linguistic gestures. Machine translation is likely to be enough to assist you in this way.

I dare to say your content, which you created with intention, deserve more dedicated resources to ensure it is correctly interpreted and understood in the context. The result should be not only semantically accurate but also accomplish your goals. You’ve crafted the contents to get your message across better to your audience, to change their minds, to touch them, to build trust or to share your creative visions. Your translation, too, should serve your goals.

Most contents are more like literature than like fortune cookies. Personally, I believe your brand should allocate a little extra resource on your written communication for a consistent brand experience. We live in the era of brands. People seek to experience rather than to consume. Each communication contributes to what your brand is and how it is perceived.

Animal Crossing: New Horizons is a Nintendo Switch game famous for its high-quality localisation from Japanese to multiple languages. I sometimes imagine how different playing Animal Crossing: New Horizons would have been if Nintendo tried to minimise the cost on their localisation. I doubt it would have been as well-loved in the English-speaking world without the top-notch translation that it has. This is a related article by translator Natalie Dawn: Localisation done right: Translating the world of Animal Crossing

Conclusion? Assess Cost & Risk

At this point you might think I am completely biassed against machine translation. I probably am; I do value quality writing and insightful interpretation of a text and AI hasn’t caught up with competent professionals in that area.

But I understand everyone’s needs vary. Everyone’s budget varies. As I mentioned earlier, it is a good idea to do your cost risk assessment before going with machine translation. And please make sure you check your tool’s terms and conditions and privacy policy carefully to prevent unknowingly giving away your unique contents and/or sensitive data to a random tech company!

photo credit: Rock'n Roll Monkey